Surface Zero — operating record
Surface Zero is a comment-only GitHub advisory agent. This page is its live operating record, read from the run ledger rather than written by hand, plus the population breakdown that keeps those numbers honest. Full build writeup: B11 — Surface Zero Goes Live.
01Live counters
Read from the run ledger. Last write 0 h ago.
02What's actually in those numbers
Every count above is drawn from one or more of six evidence classes. They are never collapsed into a single number — blending evaluation traffic, production infrastructure checks, and real external adoption into one figure would hide which kind of evidence is actually backing the claim.
| Class | Definition | Issues | Runs | Real cost |
|---|---|---|---|---|
| Organic | A naturally-occurring eligible issue opened by someone who was not generating a test case | none observed | 0 | $0.00 |
| Evaluation (batch 2) | Blind-test scenarios written by Gemini, blind to the fact that classification was being tested, scored against a pre-written key | #36–69 (34) | 34 | $0.1572 |
| Evaluation (batch 1) | Same discipline, run earlier via the offline CLI harness — a separate evidence store, not present in this ledger at all | #1–16 (16) | n/a (local JSON results) | not tracked in this ledger |
| Verification | Deliberate manual test issues created to confirm a specific fix behaves correctly in production | #70–76 (7) | 7 | $0.1313 |
| Discarded test artifact | Gemini's accidental verbatim reproduction of batch 1 as new issues, caught same day, deleted from GitHub — kept because the ledger is append-only and these runs really happened | #18–35 (18) | 18 | $0.0830 |
| Non-actionable (ignored) | Webhook deliveries that never needed processing at all (issue edits, non-issue events) — should_process returns false before any issue data is extracted | n/a | 24 | $0.00 |
Excluded from the table entirely: adamsaulters/test-smoke (6 runs, 2026-09-22), pre-App-migration pipeline validation on a disposable repo, before EverForge was the confirmed target. Not Gate A evidence of any kind.
03Uptime, honestly
This is not a measured percentage. Surface Zero has no continuous health-check or heartbeat — only a record of when webhook deliveries were actually received and processed. A gap between two run timestamps could mean the service was down, or could simply mean no GitHub event fired in that window. Those are different claims, and reporting a percentage here would fabricate precision the evidence doesn't support.
What can be said honestly: the service has processed real webhook deliveries since 2026-09-24, and the counter above reports uptime as days-since-first-activity, not days-of-confirmed-availability. No continuous-uptime claim is made. The single biggest observability gap in the current setup is that a quiet day and a down day look identical in this data.
04Cost per successful outcome, defined narrowly on purpose
“Successful outcome” is frozen, for this page, as cost per successfully evaluated recommendation: a run against a blind-test issue that reached a valid terminal state — succeeded or abstained, not failed. That's deliberately the narrowest, most defensible of three possible layers:
- successful technical run — not this metric
- successful evaluated recommendation — this metric
- human-accepted recommendation — not measurable yet; no independent maintainer exists to accept or reject anything
Current value: $0.1572 / 34 runs = $0.0046 per successfully evaluated recommendation, batch 2 only. Zero failed runs occurred in that population, so the number isn't masking error-driven cost inflation — but it's also a small, self-evaluated sample, not an independent judgment of usefulness.
05Override rate
Not yet meaningful. Override rate would measure the fraction of label suggestions an independent repository maintainer explicitly disagreed with or changed. No independent external-maintainer decision stream exists yet — Adam is both the system's author and the only person who has ever acted on an issue here, so there is no independent signal to measure. Shown here as explicitly not-yet-measurable, never omitted and never reported as a false zero.
06Incidents
The posting credential's comments attributed to Adam's own GitHub account rather than a distinct bot identity, despite the token's permission scope being correctly narrow. Scope and identity are different properties of a credential, and nothing in the methodology up to that point had asked a test to check the second one. Found by reading a live comment's byline, not by any automated check. Fixed by migrating to a GitHub App installation token; re-verified live afterward as surface-zero-advisor[bot].
Live counters read from the run ledger, last write 0 h ago. The population table is a point-in-time pull on the weekly operating cadence, not an automatic feed.